Last refresh: 2023-05-16 03:44:01.219441
The final day is held back from the model to assess performance on unseen data for the forecast.
Last data model was trained on: 2023-05-13 21:45:00.Back test prediction for: 2023-05-14Forecast from: 2023-05-16Seen Mean Absolute Error (per period): 545.01 MWThis is a tabular view of the main graph.
This is a quick and dirty experiment using Belgium’s largest grid operator (Elia)’s generous open data commitment to model the previous load and forecast ahead a 3 day horizon.
To check with the actual grid situation, visit here and scroll down to the grid load chart.
Every day at midnight UTC a script automatically downloads the previous day’s load from the elia website, as well as the temperatures in Brussels and Bruges from the openmeteo API.
A machine learning technique called gradient boosting is used to make a time series prediction for the next 3 days.
A number of different features are included in the final model, including holiday and weekend times, warming effects and recent loads.
A back fit on the past day is also run as a quick a validation step. This allows a Mean Absolute Error calculation: the average error expected, either + or - per 15 minute period.
For some exploratory data analysis I sketched up while working on this model, see here.